This dataset is released for non-commercial academic research only.
Access is granted manually after paper acceptance. Please describe
your intended use and affiliation below.

PCL-PASI

PCL-PASI (Preclinical PASI) is a multi-task multi-device image dataset for psoriasis severity assessment in mice. Each image carries three independent severity scores — Erythema, Thickness, Scaling — on a simplified 3-class scale (0 None, 1 Mild, 2 Severe). Images were captured with three devices (Sony α6500, Xiaomi 12 Lite, iPhone 15 Pro) across an 11-day experimental window.

🔒 Access policy (gated, manual approval)

Review phase: We collect access requests but do not grant access during peer review (exceptions only for co-authors or formal collaborators).
Post-acceptance: We will evaluate and approve requests individually. Priority is given to academic/non-commercial projects with a clear research purpose.
Requirements: Institutional email, brief usage statement, and agreement to the license and access-form terms.
Technical: Approved users must be authenticated (login or token) to load the dataset.

At a glance

Total images: 24,766
Subjects: 43 mice (S01 … S43)
Cameras: 3 (CAM1 = Sony α6500, CAM2 = Xiaomi 12 Lite, CAM3 = iPhone 15 Pro)
Timeframe: 13.05.2025 – 23.05.2025
Tasks: 3 correlated classification tasks (E / T / S), 3 classes each
Ethics approval: 2024/33, Karadeniz Technical University Animal Experiments Local Ethics Committee

Experimental groups

Distribution of the 43 subjects across experimental groups with sequential identifiers.

Group ID	Description	Subject IDs	Total
`HC-N`	Healthy Control (Negative)	S01 – S06	6
`HC-TCH`	Healthy + Test Compound (High Dose)	S07 – S12	6
`PM-U`	Psoriasis Model (Untreated)	S13 – S18	6
`PM-TCL`	Psoriasis Model + Test Compound (Low Dose)	S19 – S24	6
`PM-TCM`	Psoriasis Model + Test Compound (Mid Dose)	S25 – S30	6
`PM-TCH`	Psoriasis Model + Test Compound (High Dose)	S31 – S37	7
`PM-SR`	Psoriasis Model + Standard Reference	S38 – S43	6
Total			43

Quick load

Images are shipped as sharded parquet files (ImageNet-1k style, ~500 MB each) with PNG bytes embedded in the image column. Loaders fetch a few dozen files instead of one HTTP request per image, which keeps free-tier users under the 1,000-call API limit.

With 🤗 Datasets

from datasets import load_dataset
# Access must be granted via the repo's "Request access" button.
ds = load_dataset("pharmaiktu/pcl-pasi")
print(ds)                 # DatasetDict with train / val / test
print(ds["train"][0])     # {"image": <PIL.Image>, "image_id": ..., "erythema": 1, ...}

# Or load a single split directly
train = load_dataset("pharmaiktu/pcl-pasi", split="train")

Streaming (no full download)

from datasets import load_dataset
ds = load_dataset("pharmaiktu/pcl-pasi", split="train", streaming=True)
for ex in ds.take(5):
    print(ex["image_id"], ex["image"].size,
          ex["erythema"], ex["thickness"], ex["scaling"])

Label-only access with pandas

For label/metadata analysis without downloading any images, read the small metadata.parquet index (~340 KB):

import pandas as pd
df = pd.read_parquet("metadata.parquet")
# columns:
#   image_id, file_path, subject_id, group_id, group_description,
#   date, camera_id, erythema, thickness, scaling, split
train = df[df["split"] == "train"]

Directory structure

PCL-PASI/
├── data/                                  # sharded parquet (image bytes embedded)
│   ├── train-00000-of-00032.parquet       # 32 train shards, ~500 MB each
│   ├── …
│   ├── val-00000-of-00008.parquet         # 8 val shards
│   ├── …
│   └── test-00000-of-00008.parquet        # 8 test shards
├── metadata.json            # canonical per-image annotations (label-only index)
├── metadata.parquet         # same data, HF-native (+ `split` column)
├── splits.json              # official subject-level train/val/test split
├── LICENSE                  # CC BY-NC 4.0
└── README.md                # this file

Each parquet row carries the full record — the embedded image plus image_id, subject_id, group_id, group_description, date, camera_id, erythema, thickness, scaling. The file_path column in metadata.json / metadata.parquet describes the original per-PNG layout for documentation; the published dataset only contains the parquet shards.

Metadata schema

Each record in metadata.json / metadata.parquet:

{
    "image_id": "S01_13.05.2025_CAM1_0001",
    "file_path": "data/S01/13.05.2025/CAM1/0001.png",
    "subject_id": "S01",
    "group_id": "HC-N",
    "group_description": "Healthy Control (Negative)",
    "date": "13.05.2025",
    "camera_id": "CAM1",
    "scores": {
        "erythema": 0,
        "thickness": 0,
        "scaling": 0
    }
}

Label semantics

Score	Meaning
0	None — no visible symptom
1	Mild
2	Severe

The three scores are independent; each image carries one label per score.

Acquisition

ID	Device	Std. resolution (W × H)
CAM1	Sony α6500 (18–200 mm E-mount OIS)	1024 × 684
CAM2	Xiaomi 12 Lite (Main Rear 108 MP, f/1.9, 26 mm)	1024 × 576
CAM3	iPhone 15 Pro (Main Rear 48 MP, f/1.78, 24 mm)	1024 × 768

Images were downscaled to width = 1024 px preserving each camera's native aspect ratio.

Official split

A subject-level, group-stratified split is provided in splits.json (seed = 42):

Split	# Subjects	# Images	Share
Train	29	17,160	69.3%
Val	7	3,537	14.3%
Test	7	4,069	16.4%

Please use this split for any numbers you report — it is the reference split for the dataset and is also pre-computed in the split column of metadata.parquet.

Don't split by camera. Images cluster by device in pretrained feature space, and train/test splits that cross camera boundaries yield misleadingly optimistic metrics. Always split at the subject level.

License

Released under CC BY-NC 4.0. This dataset is for non-commercial academic research only.

Not for clinical use. The imiquimod mouse model does not transfer directly to human psoriasis. No diagnostic conclusions about human patients should be drawn from models trained on this data.

Citation (to be updated post peer review)

Please use the provisional citation below during peer review. A final citation with full bibliographic details will be provided after acceptance.

PharmAI KTU Research Group (2026).
PCL-PASI: Multi-Task Multi-Device Psoriasis Dataset [Dataset].
Hugging Face: https://huggingface.co/datasets/pharmaiktu/pcl-pasi

Ethics

Ethics approval 2024/33, Karadeniz Technical University Animal Experiments Local Ethics Committee. Dosing, treatment routes, and sacrifice schedule follow standard published protocols for this rodent psoriasis model. Image capture did not add an intervention beyond routine handling.

Downloads last month: 62

Total file size:

20.2 GB